Very slow Ni convergence

[svollebregt]

I have problems with the convergence of a Ni system consisting of 90 atoms configured in a two layer slab model with some vacuum space in between. I'm using the double zeta polarized basis set (also tried the double zeta double polarized set, without improvement) and tried several different values for the k point sampling (from 4, 4, 1 to 10, 10, 1 with increments of 2). I also tried extending the cut-off energy from the default 75 Hartree to 150 Hartree. Decreasing the electron temperature to 10 K gave no improvement. However, when I increased the temperature to 1000 K for the first time the system converged within 100 steps (to be precise: in 98 steps using k=661 and a cut-off of 150 Hartree). When I increased the k points to 8, 8, 1 the system didn't convergence within 100 steps, however.

Are there more settings I could try to speed up convergence? I used a similar system for Cu, Pd and Al and none of those materials had convergence issues (although Pd took about 60 steps to convergence vs ~30 for Cu and Al). I want to optimize the geometry of this system, and that is going to take ages if the system converges this slow.

I attached the configuration to my post.

[Anders Blom]

Ni is naturally spin-polarized, so I would run it as such

[svollebregt]

Is the attached script the correct way to run spin-polarized calculations? Is there a procedure in setting the initial spin?

[Anders Blom]

Check the isolated Ni atom first, that gives you an idea of the initial spin. But starting at 1. should normally work fine, it should relax quite fast. You can make the script a bit more clever by using

Code: python

initial_spin = InitialSpin(scaled_spins=[1.0,]*len(elements))

[svollebregt]

I'm currently running the above script, but after 84 steps there is still no sign of convergence.

[Anders Blom]

Try reducing the number of history steps to 4. We had some success with that for Cu, perhaps Ni has a similar feature, which is that the atomic and "configuration" electronic structure is very different. Having a long mixing history makes it harder to "forget" the atomic structure.

[svollebregt]

Unfortunately reducing the history to 4 didn't help either, would changing the damping factor help (now 0.2)? It seems from the output that the system gets close to the total energy, but keeps oscillating quite extensively around it (dE is  the order of eV and also the dH term is big, in the order of 10^1).

I will now try to see what happens when I increase the history, comparing the output to the standard history it seems that the shorter history actually worsened the convergence. At default I could see the delta terms go down, but at a sudden point increase again.

[svollebregt]

Increasing the history to 15 seems to improve, as the dE term now actually manages to reach the set limit of 1e-06. However, the dH term keeps oscillating and doesn't become smaller than the set limit.

[nori]

Please try attached.
I could get the system converged with it in 69 steps.

Please note that exchange_correlation(LSDA or SGGA) must be designated explicitly in the script in order to do spin-polarized calculation properly.

[Nordland]

I have also tried to create a Nickel 3x3 slap and it converges as well for me using just LDA.

[svollebregt]

This is strange... I ran the 3x3_slap file and it didn't converge on my system. I'm running ATK 10.8.0 on 3 systems, each 2 cores. I attached the output to this post. It seems that the dE term gets small enough, however the dH term never reaches the default accuracy.

[Anders Blom]

My suspicion is that this is related to MPI. You run your calculations in parallel, while Nordland ran in serial, am I right?

If so, a fix is both identified and implemented, and will be released later this week, hopefully.

[svollebregt]

Running a non-threaded simulation now. Will take some time but tomorrow I will know the outcome.

[Anders Blom]

Threading is not the issue, the bug is in the MPI code. So you need to run it in serial, without any mpiexec.

[svollebregt]

I ran it only on my main node, so without mpiexec, and still Nordlands configuration doesn't converge within 100 steps. The output is again attached.